Insights into dynamic sound localisation: A direction-dependent comparison between human listeners and a Bayesian model.

Abstract

Self-motion is an essential but often overlooked component of sound localisation. While the directional information of a source is implicitly contained in head-centred acoustic cues, that acoustic input needs to be continuously combined with sensorimotor information about the head orientation in order to decode these cues to a world-centred frame of reference. On top of that, the use of head movement significantly reduces ambiguities in the directional information provided by the incoming sound. In this work, we evaluate a Bayesian model that predicts dynamic sound localisation, by comparing its predictions to human performance measured in a behavioural sound-localisation experiment. Model parameters were set a-priori, based on results from various psychoacoustic and sensorimotor studies, i.e., without any post-hoc parameter fitting to behavioral results. In a spatial analysis, we evaluated the models capability to predict spatial localisation responses. Further, we investigated specific effects of the stimulus duration, the spatial prior and sizes of various model uncertainties on the predictions. The spatial analysis revealed general agreement between the predictions and the actual behaviour. The altering of the model uncertainties and stimulus duration revealed a number of interesting effects providing new insights on modelling the human integration of acoustic and sensorimotor information in a localisation task. Author summaryIn everyday life, sound localisation requires both interaural and monaural acoustic information. In addition to this, sensorimotor information about the position of the head is required to create a stable and accurate representation of our acoustic environment. Bayesian inference is an effective mathematical framework to model how humans combine information from different sources and form beliefs about the world. Here, we compare the predictions from a Bayesian model for dynamic sound localisation with data from a localisation experiment. We show that we can derive the model parameter values from previous psychoacoustic and sensorimotor experiments and that the model without any post-hoc fitting, can predict general dynamic localisation performance. Finally, the discrepancies between the modelled data and behavioural data are analysed by testing the effects of adjusting the model parameters.